KR101987475B1 - Neural network parameter optimization method, neural network computing method and apparatus thereof suitable for hardware implementation - Google Patents

Abstract

The present invention relates to a method for optimizing a neural network parameter suitable for hardware implementation, a method for operating a neural network, and an apparatus thereof. According to the present invention, the method for optimizing a neural network parameter suitable for hardware implementation comprises the steps of: performing shape transformation on an existing parameter of a neural network into a code parameter and a size parameter having a single value per a channel; and generating an optimized parameter by pruning the shape-transformed size parameter. Therefore, the accuracy loss is minimized and an operation speed can maximized by effectively optimizing a large amount of operations and parameters of a convolution neural network in hardware implementation.

Description

[0001] The present invention relates to a neural network parameter optimization method, a neural network parameter optimization method,

Field of the Invention The present invention relates to computational techniques in artificial neural networks, and more particularly, to a method of neural network parameter optimization, a neural network computation method, and apparatus therefor, which are suitable for hardware implementation.

Recently, with the development of artificial intelligence technology, artificial intelligence technology is applied and introduced in various industries.

According to this trend, there is an increasing demand for real-time hardware implementation of an artificial neural network such as a Convolutional Neural Network.

However, the artificial neural network-related arithmetic method according to the related art is not easy to implement because of many parameters and computation amount of the convolution neural network.

In order to solve this problem, there have been proposed parameter optimization methods for reducing the amount of computation compared with accuracy, such as a method of omitting unnecessary parameters through learning and a method of reducing the number of bits of parameters.

However, since the parameter optimization methods according to the prior art are limited to a software operation method that does not consider a hardware implementation, there is a limit to implementing the convolutional neural network in real time hardware by applying only this method.

Therefore, a realistic and applicable technology for artificial neural network related computation technology suitable for hardware implementation is desperately needed.

Published Patent Publication No. KR 10-2011-0027916 (published date March 17, 2011)

SUMMARY OF THE INVENTION The present invention is conceived to solve the problems described above, and it is an object of the present invention to provide a neural network parameter optimization method for optimizing a large amount of computation and parameters of a convolutional neural network to hardware implementation to obtain a minimum accuracy loss and a maximum computation speed, And to provide a neural network computing method and apparatus therefor.

In addition, the present invention is to provide a neural network parameter optimization method, a neural network operation method, and a device thereof, which can correct an accuracy through a minimum operation speed loss.

The neural network parameter optimization method according to an embodiment of the present invention includes the steps of transforming existing parameters of a neural network into a size parameter having a sign parameter and a single value per channel, .

The sign parameter determines the direction of the elements of each channel of the existing parameter, and the size parameter optimizes the weights to a single representative value per channel of the existing parameter.

In the transforming step, a magnitude parameter may be generated by calculating an average of absolute values of the elements of each channel of the existing parameter.

If the predetermined element constituting the channel of the existing parameter is less than 0, the element value of the corresponding code parameter is set to 0, and if it is equal to or larger than 0, the element value of the corresponding code parameter is represented by 1 Code parameters can be generated.

In the pruning step, a reference value is calculated using an average value and a size distribution of size parameters for each input channel or an average value and a size distribution of size parameters for input and output channels, and a reference value having a value smaller than the calculated reference value The size parameter value may be set to 0 to omit the convolution operation of the corresponding channel.

When a pruning is performed using an average value and a size distribution of size parameters of input channels constituting a predetermined layer, a reference value is calculated by multiplying an average value of size parameters of input channels by a constant for each layer, If the value of the size parameter of the channel is smaller than the reference value, the size parameter value of the channel is changed to 0 and pruned. By using the average value and the size distribution of the size parameters of the input and output channels constituting the predetermined layer, The reference value is calculated by multiplying the average value of the size parameters of the input channels and the output channels by a value obtained by multiplying the average value of the size parameters of the input channels and the output channels by a layer specific constant. If the size parameter value of the predetermined input channel is smaller than the reference value, 0 to be pruned.

The constant value for each layer may be determined according to the distribution of convolution parameters for each layer.

The neural network parameter optimization method may further include a step of transforming the existing parameters of the neural network into rate parameters.

The step of transforming into the ratio parameter may comprise the steps of variably allocating bits of the ratio parameter and quantizing the range and the weight of the ratio parameter element by user selection.

The neural network operation method according to another embodiment of the present invention includes the steps of loading an existing parameter and input channel data of a neural network into a memory and converting the existing parameter into a size parameter having a single value per code parameter and channel, Convolutionalizing the optimized parameters and input channel data, correcting the optimized parameters, and adjusting the existing parameters to the corrected optimized parameters And updating it.

The neural network computation method includes the steps of: determining whether a learned parameter exists; generating an initial parameter through parameter initialization if the learned parameter is not present; generating an optimized parameter for the initial parameter; And if the parameter exists, loading the existing parameter.

The step of convoluting and optimizing the optimized parameter and the input channel data includes the steps of loading optimized parameters into a memory and determining whether the value of the magnitude parameter included in the optimized parameter loaded in the memory is 0 And omitting the convolution operation if the value of the magnitude parameter is zero.

The step of convoluting and calculating the optimized parameter and the input channel data comprises the steps of determining the direction of the data by bitwise computing the sign parameter and the input channel data if the value of the magnitude parameter is not 0, Calculating a sum of the input channel data by the size of the input channel data; and performing a single multiplication operation on the size parameter and the input channel data.

The step of convoluting and calculating the optimized parameter and the input channel data may further include the step of reducing the error of the calculation result by differentiating the weight parameter in the size parameter using the ratio parameter when the ratio parameter is present .

According to another aspect of the present invention, there is provided a neural network computing apparatus including a code parameter converting unit for converting an existing parameter into a code parameter, a size parameter converting unit for converting an existing parameter into a magnitude parameter having a single value per channel, And a parameter pruning unit for pruning the transformed size parameter to generate an optimized parameter.

The parameter pruning unit calculates a reference value using an average value and a size distribution of size parameters for each input channel or an average value and a size distribution of size parameters for input and output channels and calculates a reference value having a value smaller than the calculated reference value The parameter value may be set to 0 to omit the convolution operation of the corresponding channel.

When the pruning is performed using the average value and the size distribution of the size parameters of the input channels constituting a predetermined layer, the parameter pruning unit multiplies the average value of the input channel size parameters by a constant for each layer, If the size parameter value of a predetermined input channel is smaller than the reference value, the size parameter value of the corresponding channel is changed to 0 and pruned. The average value and size of the size parameters of the input and output channels constituting the predetermined layer When the pruning is performed using the distribution, a reference value is calculated by multiplying the average value of the size parameters of the input channels and the output channels by a constant for each layer. If the size parameter value of the predetermined input channel is smaller than the reference value, It can be pruned by changing the size parameter value of the channel to zero.

The neural network computation apparatus may further include a rate parameter conversion unit for converting the existing parameters into rate parameters.

The neural network computing apparatus may further include a reasoning unit for convoluting and optimizing the optimized parameters and the input channel data.

The reasoning unit may determine whether the value of the magnitude parameter included in the optimization parameter is 0, and omit the convolution operation if the value of the magnitude parameter is zero.

If the ratio parameter is present, the reasoning unit may reduce the error of the operation result by differently reflecting the weight value in the size parameter using the ratio parameter.

As described above, the present invention provides a neural network parameter optimization method, a neural network operation method, and a device thereof, which can optimize a large amount of computation and parameters of a convolutional neural network to hardware implementations to obtain a minimum accuracy loss and a maximum operation speed .

Further, according to the present invention, by optimizing the parameters of the neural network, it is possible to satisfy the low power and high performance required in the hardware implementing the convolution neural network operating in real time. If the neural network parameter optimization technique according to one embodiment is applied, computation can be omitted for each channel according to the size of the channel in the convolution operation. In addition, the hardware does not determine whether to omit or omit the operation for each element in the channel, but it determines whether to omit operation for each channel, so that the operation is reduced by a multiple of the number of channel elements.

Further, the present invention can correct the accuracy through a minimum operation speed loss. For example, the scale parameters can be separated to effectively apply different weights according to the range of the values, thereby effectively improving the performance in hardware implementation.

1 is a diagram illustrating a convolution parameter used in a neural network convolution operation according to an embodiment of the present invention;
FIG. 2 illustrates a learning process for generating optimization parameters in a Convolution Neural Network (CNN) according to an embodiment of the present invention; FIG.
FIG. 3 is a detailed view illustrating a learning process for generating optimization parameters of FIG. 2 according to an embodiment of the present invention. FIG.
FIG. 4 is a detailed view of the reasoning process according to an embodiment of the present invention. FIG.
5 is a diagram illustrating an example of application of a convolution parameter optimization process according to an embodiment of the present invention;
FIG. 6 illustrates a comparison between a general convolution operation and a convolution operation according to an embodiment of the present invention;
Figure 7 illustrates the benefits of convolution operations after parameter optimization in accordance with an embodiment of the present invention;
8 is a diagram illustrating an example of application of a convolution parameter optimization process according to another embodiment of the present invention;
9 is a view comparing a parameter distribution applying only a magnitude parameter and a parameter distribution added a ratio parameter in order to explain a ratio parameter optimization;
10 illustrates an example of a convolution operation using an optimized parameter according to an embodiment of the present invention;
11 is a diagram illustrating the configuration of a neural network computing apparatus according to an embodiment of the present invention.
FIG. 12 is a detailed block diagram of the parameter optimizer of FIG. 11 according to an embodiment of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention and the manner of achieving them will become apparent with reference to the embodiments described in detail below with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the invention to those skilled in the art. Is provided to fully convey the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS In the following description of the present invention, detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear. , Which may vary depending on the intention or custom of the user, the operator, and the like. Therefore, the definition should be based on the contents throughout this specification.

Each block of the accompanying block diagrams and combinations of steps of the flowcharts may be performed by computer program instructions (execution engines), which may be stored in a general-purpose computer, special purpose computer, or other processor of a programmable data processing apparatus The instructions that are executed through the processor of the computer or other programmable data processing apparatus will produce means for performing the functions described in each block or flowchart of the block diagram.

These computer program instructions may be stored in a computer usable or computer readable memory capable of directing a computer or other programmable data processing apparatus to implement a function in a particular manner, It is also possible for the instructions stored in the block diagram to produce an article of manufacture containing instruction means for performing the functions described in each block or flowchart of the flowchart.

And computer program instructions may also be stored on a computer or other programmable data processing apparatus so that a series of operating steps may be performed on a computer or other programmable data processing apparatus to create a computer- It is also possible for the instructions to perform the data processing apparatus to provide steps for executing the functions described in each block of the block diagram and at each step of the flowchart.

Also, each block or step may represent a portion of a module, segment, or code that includes one or more executable instructions for executing the specified logical functions, and in some alternative embodiments, It should be noted that functions may occur out of order. For example, two successive blocks or steps may actually be performed substantially concurrently, and it is also possible that the blocks or steps are performed in the reverse order of the function as needed.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the following embodiments of the present invention may be modified into various other forms, and the scope of the present invention is not limited to the embodiments described below. The embodiments of the present invention are provided to enable those skilled in the art to more fully understand the present invention.

FIG. 1 is a diagram illustrating a convolution parameter used in a neural network convolution operation according to an exemplary embodiment of the present invention. Referring to FIG.

1, when the total number of convolutional layers of a neural network is denoted by l, k output channels (O ₁ , ..., O _k ) exist for each layer, The output channel consists of j Input Channels (I ₁ , I ₂ , ..., O _j ). Each input channel has i weight parameters W ₁ , W ₂ , ..., W _i . In the present invention, the optimization technique is formally described based on the definition of the convolution parameter defined above.

2 is a diagram illustrating a learning process for generating optimization parameters in a Convolution Neural Network (CNN) according to an embodiment of the present invention.

2, the learning process may include an existing parameter and input data loading step S200, a parameter optimization step S210, an inference step S220, a correction step S230, and a parameter update step S240 .

More specifically, when the parameter optimization step (S210) is omitted, it is the same as a general deep learning learning procedure, which means that the present invention can be easily applied in a general deep learning learning environment.

The convolution neural network optimization technique according to the embodiment of the present invention can perform the convolutional neural network optimization method by adding only the parameter optimization step S210 to the existing learning environment when the existing learned parameters exist. At this time, if the existing learned parameter does not exist, the parameter optimization step (S210) may be applied after the initial learning is performed through learning omitting the parameter optimization step (S210). On the other hand, when the learning is performed by applying the parameter optimization step (S210) from the beginning, the result accuracy may be lower than the same calculation decrease amount. Therefore, the existing learned convolution parameter and input data are loaded into the memory (S200) The convolution parameter optimization method proposed by the present invention is applied to the loaded existing convolution parameter (S210).

Next, the inference (S220) is performed by convoluting the parameter to which the optimization is applied and the input data, and the existing parameter is updated through the parameter correction step (S230) (S240). In the parameter correction step (S230), a process of obtaining back propagation is included. After the learning ends, the convolution parameter optimization technique is applied to store the parameters. The parameter optimization step (S210), the inference step (S220), the parameter correction step (S230), and the update step (S240) may be repeatedly performed until the target number of times is reached.

FIG. 3 is a detailed view illustrating a learning process for generating optimization parameters of FIG. 2 according to an embodiment of the present invention.

Referring to FIG. 3, the neural network computing apparatus according to the embodiment of the present invention determines whether there is a learned parameter (S300). Next, if the learned parameter exists, it is loaded into the memory (S302). Subsequently, it is determined whether or not to apply the convolution parameter optimization S303 of the present invention (S304). The case where the convolution parameter optimization (S303) is not applied is the same as the general deep learning learning process.

The convolution parameter optimization step (S303) includes a size parameter generation step (S306), a sign parameter generation step (S308), and it is determined whether to apply ratio parameter optimization (S310) (S312) of generating a ratio parameter when it is determined that the ratio parameter is not satisfied.

According to the embodiment of the present invention, when the convolution parameter optimization step (S303) is completed, the neural network computation apparatus performs inference with parameters and input data to which the optimization is applied through inference step (S316) And updates the existing parameter with optimized and corrected parameters (S320). At this time, it is determined whether the target number of times has been reached (S322). If not, optimization parameter optimization step (S303), inference step (S316), correction step (S318) and parameter update step (S320) are repeated.

4 is a detailed view illustrating a reasoning process according to an embodiment of the present invention.

Referring to FIG. 4, the neural network computation apparatus according to an embodiment of the present invention loads optimization parameters into a memory (S400) and determines whether or not a size parameter value exists, that is, whether a size parameter value is zero (S410). If the size parameter value is 0, the inference operation amount can be reduced since the subsequent process is omitted.

Next, if the size parameter value is not 0, the input data is loaded into the memory (S420), and the sign parameter and the input data are bit-operated (S430). Then, a sum operation is performed between the input data (S460) and a multiplication operation of the size parameter and the input data (S470) is performed. After the bit operation (S430), if the ratio parameter is applied (S440), referring to the ratio parameter, applying the ratio parameter corresponding constant to the size parameter (S470) may be further included.

5 is a diagram illustrating an application example of a convolution parameter optimization process according to an embodiment of the present invention.

Referring to FIG. 5, the parameter optimization process according to an exemplary embodiment of the present invention is mainly composed of a shape conversion step 510 and a pruning step 520. The parameter type conversion 510 refers to a process of reducing the amount of computation of the existing convolution parameter and converting it into a form that is easy to prune. For example, as shown in FIG. 5, the existing parameter 50 is divided into a magnitude parameter 52 indicating a magnitude and a sign parameter 54 indicating a direction. FIG. 5 shows an example in which the existing parameters have three input channels 50-1, 50-2, and 50-3. Each of the input channels 50-1, 50-2, and 50-3 includes a plurality of parameter elements. The sign parameter 54 determines the direction of the elements of the channel of the existing parameter 50 and the magnitude parameter 52 optimizes the weights to a single representative value per channel of the existing parameter 50.

The user can arbitrarily set the number of bits of each constituent element of the size parameter 52, and can be represented by an integer and a floating. As the number of bits increases, the same result as the original convolution operation result is obtained. The number of bits of each constituent element in the sign parameter 54 is 1 bit, which indicates only positive (+) direction and negative (-) direction. Various methods can be applied to separate the magnitude parameter 52 and the sign parameter 54 from the existing parameter 50. For example, in the conventional parameter 50, the type conversion can be performed through the average of the input channels of the convolution layer and the most significant bits of each parameter.

One of the form conversion methods to the size parameter 52 is to use the average value of the constituent elements (weight parameters) of the input channel. For example, it is assumed that there are i constituent elements W _i , j input channels I _j , k output channels O _k , and size parameters M = {M ₁ , M ₂ , ... , M _j }, M _j is expressed by Equation (1).

One of the form conversion methods into the sign parameter 54 is a method of allocating a value of 0 if the constituent elements (weight parameters) of the channel are equal to or greater than 0, and assigning a value of 1 when the constituent elements (weight parameters) That is, the sign parameter S = {S ₁ , S ₂ , ... , S _i }, S _i is expressed by Equation (2).

After the morphing process is performed, a pruning step 520 is performed to remove unnecessary parameters. In the pruning step 520, low-importance parameters are selected and their values are set to 0 (zero), which makes it possible to omit calculation, thereby reducing the total amount of computation. At this time, the optimized parameter is generated by trimming the size parameter according to the size. For example, if the predetermined size parameter value is smaller than a preset reference value, the size parameter of the corresponding channel is truncated.

It is important to note that this pruning process is not applied to the entire convolution layer parameters but only to the transformed size parameters 52. [ By applying the scale parameter 520 only to the size parameter 52, the omitted parameters can affect all parameters of each input channel of the convolution layer, maximizing the effect of the pruning 520.

Various methods can be applied to omit unnecessary parameters. For example, pruning is performed using the convolution layer input channel and output channel average value and distribution. Each layer has a user-defined constant, which is determined by the distribution of convolution parameters per layer. As a reference value for omitting the value, an average value of input channel size parameters or an average of input channel and output channel size parameters may be used. The layer constant has a value of 0 to 1. When the convolution parameter per layer is assumed to be a continuous uniform distribution or a normal distribution, the layer constant may be 0.5 to 0.7 in order to omit the 50% size parameter. This value is referred to, and the layer constant is determined by considering the distribution of actual parameters.

M _LOI = {M ₁₁₁ , M ₁₁₂ , ... , M _lkj }, L = {C ₁ , C ₂ , ... , C _l ), and C is a layer constant, the size parameter pruning method applying the criterion for each input channel is expressed by Equation 3 below.

The output parameter is applied to each output channel, but the size parameter pruning method is expressed by Equation (4).

Hereinafter, examples of convolution parameter optimization calculations using Equations 1 to 4 described above will be described.

it is assumed that a size parameter and a sign parameter are calculated with respect to an existing convolution parameter having i = 4, j = 2, k = 2, and pruning is performed on the size parameter. At this _{time, O 1 = {I 11,} I 12} = {[W 111, W 112, W 113, W 114}, [W 121, W 122, W 123, W 124]]} = {[-0.5, 1.0 , 0.5, -1.0], [1.0 , -1.5, 1.0, -1.5]}, O 2 = {I 21, I 22} = {[W 211, W 212, W 213, W 214}, [W 231, W ₂₂₂ , W ₂₂₃ , W ₂₂₄ ]]} = {[1.5, -2.0, -1.5, 2.0], [-2.0, -2.5, 2.0, 2.5]}.

The magnitude parameters M = {M ₁₁ , M ₁₂ , M ₂₁ , M ₂₂ } = {0.75, 1.25, 1.75, 2.25} calculated by applying Equation (1). That is, M ₁₁ = 1/4 (0.5 + 1.0 + 0.5 + 1.0) = 0.75, M ₁₂ = 1/4 (1.0 + 1.5 + 1.0 + 1.5) = 1.25, M ₂₁ = 1/4 + 2.0) = 1.75 and M ₂₂ = 1/4 (2.0 + 2.5 + 2.0 + 2.5) = 2.25.

Applying equation (2) a code parameter S = calculated by _{{S 1, S 2, S} 3, S 4} = {[1, 0, 0, 1], [0, 1, 0, 1], [0, 1, 1, 0], [1, 1, 0, 0]}.

Here is an example of a pruning application for the size parameter. M ₁₁₁ , M ₁₁₂ , M ₁₂₁ , M ₁₂₂ } = {0.75, 1.25, 1.75, 2.25} separated from the existing parameters for the existing parameters with j = 2, k = 2, , And the layer constant L ₁ = 0.9. When applied to the input channels of Equation (3), the pruned magnitude parameters M '= {M ₁₁₁ , M ₁₁₂ , M ₁₂₁ , M ₁₂₂ } = {0, 1.25, }to be. For example, in the case of M ₁₁₂ , since 1.25 < 1/2 (0.75 + 1.25) 0.9, M ₁₁₂ = 0 according to Equation (3).

If applied to the input channel mill output channels of Equation 4, the pruned magnitude parameters M '= {M ₁₁₁ , M ₁₁₂ , M ₁₂₁ , M ₁₂₂ } = {0, 0, 1.75, 2.25}. In the case of ₁₁₂ M, 1.25 because <1/4 (0.75 + 1.25 + 1.75 + 2.25), 0.9, and M ₁₁₂ = 0 by the equation (4). It can be determined according to the input channel or the distribution depending on the input and output channels.

FIG. 6 illustrates a comparison between a general convolution operation and a convolution operation according to an embodiment of the present invention.

Referring to FIG. 6, the general convolution operation 60 includes a multiplication operation 600 and a sum operation 602 as many as the convolution parameter filter size. However, the convolution operation 62 using the parameter optimization according to the embodiment directs the data through the bit operation 620, performs a sum operation 622 by the filter size, and then performs a single multiplication operation 624, The results can be derived. For example, as shown in FIG. 6, a general convolution operation 60 may be performed such that the multiplication operation 600 is performed nine times, but the convolution operation 62 using parameter optimization according to one embodiment is only one . This means that there is a gain in the product of the filter size.

In addition, if a convolution parameter is not transformed and a pruning is applied, it is necessary to perform a comparison operation as much as the filter size in order to determine whether or not the operation should be omitted. A typical convolution operation (60) case of the example of FIG. 6 requires a total of nine comparison operations. However, in the case of the convolution operation 62 to which the parameter optimization technique of the present invention is applied, it is possible to determine whether or not to omit the operation by only one comparison per convolution layer input channel. This reduction in the amount of computation requires fewer operators than the conventional convolution operation 60. Therefore, hardware can be efficiently designed.

Figure 7 illustrates the benefits of convolution operations after parameter optimization in accordance with an embodiment of the present invention.

Referring to FIG. 7, output data is generated by convoluting input data, size parameter, and sign parameter. When the size parameter value is 0, there is no need to perform an operation, and the amount of operation can be greatly reduced on a channel-by-channel basis.

8 is a diagram showing an application example of a convolution parameter optimization process according to another embodiment of the present invention.

The goal of convolutional parameter optimization is to achieve maximum accuracy with maximum optimization (reduction in computation). Decrease in computation and accuracy are universally inversely related. There are performance (speed and accuracy) to be satisfied in actual application level, but it is not easy for users to control it. In order to solve this problem, the present invention proposes an optimization method using a scale parameter. This is the ratio parameter added to the above-described optimization method with reference to FIG.

The shape of the existing parameter 50 is converted into the size parameter 52, the sign parameter 54 and the ratio parameter 58 in the shape conversion process 510 and the pruning 520 is applied only to the size parameter 52 do. For inference, we apply a weight to the size parameter by referring to the ratio parameter (58) to reduce the error of the existing operation result.

The ratio parameter 58 indicates how important the magnitude parameter 52 is. In other words, we compare the importance of the convolution layer input channel parameters and apply a constant difference. For one constituent element of the ratio parameter 58, the user can assign the desired bit and the user can thereby adjust the degree of optimization. The ratio parameter 58 does not imply a weight, and there are constant values that refer to the ratio parameter 58. These values are assigned values that are easy to be implemented in hardware, and can be implemented only by bit operations. If the optimization is performed by applying the weight based on the ratio parameter 58 without immediately applying the value of the size parameter 52, the degree of similarity with the existing pre-optimization parameter is increased and the accuracy is increased.

FIG. 9 is a diagram comparing a parameter distribution applying only a magnitude parameter and a parameter distribution adding a ratio parameter to explain a ratio parameter optimization.

8 and 9, various methods can be applied to generate the ratio parameter 58, in which the ratio parameter 58 is generated through the distribution of the convolution parameters. Assuming that the number of bits of the rate parameter is b, there are 2 ^b user defined arbitrary constants, which are called rate parameter corresponding constants. The ratio parameter correspondence constant can be arbitrarily set by the user based on the convolution parameter distribution, and is a constant corresponding to the ratio parameter.

An example of calculating the ratio parameter 58 is as follows.

Rate parameter Number of bits = b,

? = {? ₁ ,? ₂ , ... ,? _t }, t = {1, 2, ... , 2 ^b -1}, λ _t > λ _{t + 1} , and the ratio parameter SC = {SC ₁ , SC ₂ , ... , SC _i }, SC _i is expressed by the following equation (5).

At the time of inference, the value obtained by multiplying the ratio parameter correspondence constant to the size parameter is used as the actual calculation value.

Hereinafter, an example of the ratio parameter optimization calculation using the above-described equation (5) will be described.

It is assumed that the ratio parameter is calculated for an existing convolution parameter i = 4, j = 2, k = 2. It is assumed that the ratio parameter bit number b = 2, t = {1, 2, 3}, λ = {0.9, 0.7, 0.5}. At this _{time, O 1 = {I 11,} I 12} = {W 111, W 112, W 113, W 114, W 121, W 122, W 123, W 124} = {-0.5, 1.0, 0.5, -1.0, 1.0, -1.5, 1.0, -1.5} , O 2 = {I 21, I 22} = {W 211, W 212, W 213, W 214, W 231, W 222, W 223, W 224} = {1.5 , -2.0, -1.5, 2.0, -2.0, -2.5, 2.0, 2.5}.

According to equation (5), SC ₁₁₁ = 2. 1/4 (0.5 + 1.0 + 0.5 + 1.0) 0.7> | -0.5 | > 1/4 (0.5 + 1.0 + 0.5 + 1.0) 0.5. The ratio parameters calculated in this way _{SC i = (SC 111, SC} 112, SC 113, SC 114, SC 121, SC 122, SC 123, SC 124, SC 211, SC 212, SC 213, SC 214, SC 221, SC is _{222, SC 223, SC 224}} = {2, 0, 2, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0}.

10 is a diagram illustrating an example of a convolution operation using an optimized parameter according to an embodiment of the present invention.

As shown in the figure, there exists a magnitude parameter M, a sign parameter S, a ratio parameter SC, and a ratio parameter correspondence constant C, and a convolution when the input data In is input The calculation is as shown in Fig. In this case, the input data In = {1.2, -2.0, 0, 0, 1, 1}, M = {0.75}, S = {0,0,1,1} , 0.5, 3.1}.

In the example referring to FIG. 10, the ratio parameter correspondence constant C is set to a multiple of 2, so that a result can be derived by a shift bit operation rather than a general arithmetic operation. In this case, it has an advantage in hardware implementation.

11 is a diagram illustrating the configuration of a neural network computing apparatus according to an embodiment of the present invention.

As shown in the figure, a neural network computing apparatus 1 according to an embodiment of the present invention may include an input unit 10, a processor 12, a memory 14 and an output unit 16. [

The input unit 10 receives data including a user command, and the input data is stored in the memory 18. The processor 12 may perform a neural network operation and store the result of the operation in the memory 18 and output it via the output 16. The memory 14 may be loaded with input data and parameters. The processor 12 according to one embodiment includes a parameter optimizer 120, a speculator 122, a calibration unit 124, and an update unit 126. [

The parameter optimizing unit 120 transforms the existing parameters and truncates the size parameters among the transformed parameters. Existing parameters may already be learned parameters and loaded into memory for computation.

The parameter optimizer 120 according to the embodiment of the present invention transforms existing parameters into size parameters and sign parameters or converts them into size parameters, sign parameters, and ratio parameters. The sign parameter determines the direction of the elements of each channel of the existing parameter and the size parameter optimizes the weights to a single representative value per channel of the existing parameter. The ratio parameter indicates how important the magnitude parameter is, and the differential weight can be applied to the magnitude parameter by referring to the ratio parameter during inference.

In an embodiment of the present invention, the pruning is performed only on size parameters among the transformed parameters. Pruning means selecting low-importance size parameters and making their values zero (0). The pruned parameters can omit the convolution operation on a channel-by-channel basis, thereby reducing the overall amount of computation. Importantly, this pruning process will affect all the parameter elements that make up each input channel of the convolution layer, maximizing the effect of pruning.

The reasoning unit 122 performs convolution operation on the optimized parameters and input channel data to deduce. Here, the reasoning unit 122 may determine whether the value of the size parameter included in the optimization parameter is 0, and omit the convolution operation if the value of the size parameter is 0. In addition, if the ratio parameter is present, the inference unit 122 may reduce the error of the calculation result as the weight parameter is reflected in the size parameter by using the ratio parameter. In the embodiment of the present invention, the correction unit 124 corrects the optimized parameter, and the update unit 126 updates the existing parameter with the corrected optimization parameter.

FIG. 12 is a detailed block diagram of the parameter optimizer of FIG. 11 according to an embodiment of the present invention.

11 and 12, the parameter optimizing unit 120 according to the embodiment of the present invention includes a size parameter converting unit 1200, a sign parameter converting unit 1202, a parameter pruning unit 1204, 1206 &lt; / RTI &gt;

The code parameter conversion unit 1202 converts the existing parameters into code parameters. The size parameter converting unit 1200 transforms existing parameters into size parameters having a single value per channel. The rate parameter conversion unit 1206 converts an existing parameter into a rate parameter.

In addition, the parameter pruning unit 1204 truncates the transformed size parameter to generate an optimized parameter. For example, if the predetermined size parameter value is smaller than a preset reference value, the size parameter of the corresponding channel is truncated.

The parameter pruning unit 1204 according to the exemplary embodiment of the present invention calculates a reference value using an average value and a size distribution of size parameters for each input channel or an average value and a size distribution of size parameters for input and output channels, The size parameter value having a value smaller than the reference value is made 0, and the convolution operation of the corresponding channel is omitted.

In case of pruning using the average value and size distribution of the size parameters of the input channels constituting the predetermined layer, the parameter pruning unit 1204 multiplies the average value of the input channel size parameters by a constant for each layer And if the size parameter value of a predetermined input channel is smaller than the reference value, the size parameter value of the corresponding channel can be changed to 0 and pruned.

When the pruning unit 1204 pruning the average value and the size distribution of the size parameters of the input and output channels constituting the predetermined layer, the parameter pruning unit 1204 calculates the average value of the size parameters of the input channels and the output channels And if the size parameter value of a predetermined input channel is smaller than the reference value, the size parameter value of the corresponding channel can be changed to 0 and pruned.

The embodiments of the present invention have been described above. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the present invention is defined by the appended claims rather than by the foregoing description, and all differences within the scope of equivalents thereof should be construed as being included in the present invention.

10: Input unit
12: Processor
14: Memory
16: Output section
120: parameter optimization section
122: Reasoning section
124:
126:
1200: size parameter conversion unit
1202: Code parameter conversion unit
1204: Parameter pruning section
1206: Rate parameter conversion section

Claims (21)

Transforming existing parameters of the neural network into a code parameter and a size parameter having a single value per channel using a code parameter conversion unit and a size parameter conversion unit constituting the parameter optimization unit; And
And pruning the transformed size parameter using a parameter pruning unit constituting the parameter optimizing unit to generate an optimized parameter,
The pruning using the parameter pruning unit may include calculating a reference value using an average value and a size distribution of size parameters for input channels or an average value and a size distribution of size parameters for input and output channels, The size parameter value having a value smaller than the value is set to 0 to omit the convolution operation of the corresponding channel,
In the pruning step using the parameter pruning unit, when pruning is performed using the average value and the size distribution of the size parameters of the input channels constituting a predetermined layer, a constant per layer is added to the average value of the size parameters of the input channels And when the size parameter value of the predetermined input channel is smaller than the reference value, the size parameter value of the channel is changed to 0 and pruned,
In the pruning step using the parameter pruning unit, when pruning is performed using the average value and the size distribution of the size parameters of the input and output channels constituting the predetermined layer, the average of the size parameters of the input channels and output channels Value is multiplied by a layer-specific constant, and when the size parameter value of a predetermined input channel is smaller than a reference value, the size parameter value of the channel is changed to 0 and pruned. Way.
The method according to claim 1,
The code parameters determine the direction of the channel-specific elements of the existing parameters,
Wherein the size parameter optimizes weights to a single representative value per channel of an existing parameter.
2. The method according to claim 1,
And a size parameter is generated by calculating an average of absolute values of elements of each channel of the existing parameters.
2. The method according to claim 1,
When the predetermined element constituting the channel of the existing parameter is smaller than 0, the element value of the corresponding code parameter is set to 0 and when the element value of the corresponding code parameter is equal to or larger than 0, Wherein the neural network parameter optimization method comprises:
delete delete The method according to claim 1,
Wherein the constant value per layer is determined according to a distribution of convolution parameters per layer.
2. The method of claim 1,
And converting the existing parameters of the neural network into rate parameters using the rate parameter converting unit that constitutes the parameter optimizing unit.
9. The method of claim 8, wherein transforming the existing parameters of the neural network into rate parameters using the ratio parameter transformer comprises:
Variably allocating bits of the ratio parameter; And
And quantizing the range and the weight of the value of the ratio parameter element by user selection.
Loading existing parameters and input channel data of a neural network into a memory using a processor configuring the neural network computing device;
Converting the existing parameters of the neural network into a code parameter and a size parameter having a single value per channel by using a code parameter conversion unit and a size parameter conversion unit constituting a parameter optimization unit included in the processor, Pratecting the transformed size parameter using a stroke unit to generate an optimized parameter;
Performing convolution operation on the optimized parameter and input channel data by using an inference unit included in the processor;
Correcting the optimized parameter using a correction unit included in the processor; And
Updating an existing parameter with a corrected optimization parameter using an update unit included in the processor,
The step of convoluting and optimizing the optimized parameter and the input channel data using the inference unit may include: loading the optimized parameter into a memory; Determining whether a value of a size parameter included in the loaded optimized parameter is 0; And omitting the convolution operation if the value of the magnitude parameter is zero,
The step of convoluting and optimizing the optimized parameter and the input channel data using the inferring unit includes the steps of: determining the direction of the data by performing a bit operation on the code parameter and the input channel data if the value of the size parameter is not 0; Summing input channel data by a convolution parameter filter size; And performing a single product operation on the size parameter and the input channel data.
11. The method of claim 10,
Determining whether a learned parameter exists;
Generating an initial parameter through parameter initialization if the learned parameter is not present;
Generating an optimized parameter for an initial parameter; And
Loading the learned parameters if they exist;
Further comprising the steps of:
delete delete 11. The method of claim 10, wherein convoluting and inferring the optimized parameters and input channel data comprises:
If the ratio parameter is present, reducing the error of the calculation result by differently reflecting the weight value in the size parameter using the ratio parameter of the ratio parameter conversion unit constituting the parameter optimization unit;
Further comprising the steps of:
A code parameter conversion unit for converting the existing parameters into code parameters;
A size parameter converting unit for converting the existing parameter into a size parameter having a single value per channel; And
And a parameter pruning section for pruning the transformed size parameter to generate an optimized parameter,
The parameter pruning unit calculates a reference value using an average value and a size distribution of size parameters for each input channel or an average value and a size distribution of size parameters for input and output channels and calculates a reference value having a value smaller than the calculated reference value The parameter value is set to 0 to omit the convolution operation of the corresponding channel,
The parameter pruning unit includes:
When a pruning is performed using an average value and a size distribution of size parameters of input channels constituting a predetermined layer, a reference value is calculated by multiplying an average value of size parameters of input channels by a constant for each layer, If the size parameter value of the channel is smaller than the reference value, the size parameter value of the corresponding channel is changed to 0,
In case of pruning using the average value and the size distribution of the size parameters of the input and output channels constituting a predetermined layer, the reference value is obtained by multiplying the average value of the size parameters of the input channels and output channels by a constant for each layer And when the size parameter value of the predetermined input channel is smaller than the reference value, the size parameter value of the corresponding channel is changed to 0 and pruned.
delete delete 16. The apparatus of claim 15, wherein the neural network computing device
A ratio parameter conversion unit for converting the existing parameter into a ratio parameter;
Wherein the neural network computation apparatus further comprises:
16. The apparatus of claim 15, wherein the neural network computing device
A reasoning unit for convoluting and optimizing optimized parameters and input channel data;
Wherein the neural network computation apparatus further comprises:
20. The apparatus of claim 19, wherein the reasoning unit
Determining whether the value of the size parameter included in the optimization parameter is 0, and omitting the convolution operation if the value of the size parameter is 0.
20. The apparatus of claim 19, wherein the reasoning unit
Wherein when the ratio parameter is present, the weight parameter is reflected to the size parameter using the ratio parameter to reduce the error of the calculation result.

Images